A Novel Adversarial Approach for EEG Dataset Refinement: Enhancing Generalization Through Proximity-to-Boundary...

Area of Science:

• Neuroscience
• Machine Learning
• Signal Processing

Background:

• Deep learning (DL) models interpret electroencephalography (EEG) signals for user intention recognition.
• Generalization ability of DL models on EEG data is limited by noisy samples.
• Existing methods for noise reduction often require complex hyperparameter tuning.

Purpose of the Study:

• To propose a novel DL training framework to enhance EEG model generalization.
• To reduce the influence of noisy samples without additional hyperparameter optimizations.
• To improve the accuracy and efficiency of EEG signal interpretation.

Main Methods:

• Developed a proximity-to-boundary score (PBS) to quantify data closeness to the decision boundary.
• Integrated PBS into the DL training framework to mitigate noisy sample influence.
• Evaluated the framework on motor imagery and sleep stage EEG datasets.

Main Results:

• The proposed framework significantly improved model generalization across datasets.
• Performance gains ranged from 1.43% to 6.66% on motor imagery tasks.
• Performance gains ranged from 0.72% to 2.85% on sleep stage classification.
• Low PBS scores correlated with noisy samples that degrade training.

Conclusions:

• The novel training framework effectively reduces the impact of noisy EEG data.
• The proximity-to-boundary score accurately identifies noise influencing model training.
• This approach enhances DL model generalization for EEG signal analysis efficiently and accurately.